Revisiting Fast and Accurate RGB-D Odometry for Real-World Use by Embracing Simplicity

Robust and accurate pose estimation of a robot is essential for many tasks. Over the past decades, researchers have explored odometry estimation using various sensors, including RGB cameras, LiDARs, and RGB-D cameras. RGB-D cameras, mainly, are attractive sensors as they provide color and depth. Despite the progress made in the RGB-D odometry approaches and systems sharing code, their practical use is often non-trivial due to technical debt in several open-source implementations. In some cases, approaches have been tuned to specific benchmarks, complicating the extension of implementations to different RGB-D sensors and environments. This paper introduces an easy-to-use and easy-to-understand yet robust RGB-D odometry pipeline adaptable across sensor platforms. We focus on simplicity and contribute a non-black-box RGB-D odometry approach and proper implementation to the community. Our open-source system utilizes ORB features for correspondence estimation and employs a frame-to-map registration strategy to estimate camera poses. Our system achieves performance on par with state-of-the-art methods while running faster than the sensor frame rate on a single-core CPU. Importantly, our approach does not require integrating IMU data or additional hardware like GPUs, making it easily deployable on various platforms.